AI_Parking_Presentation

1.

P
AI Parking Management System
A cognitive prototype: perception, attention, memory and knowledge representation
working together
35,717
30
4
10
real sensor readings
Birmingham car parks
cognitive modules
if-then rules
Practical Assignment 3 · Group: ________________ · Data: UCI Parking Birmingham (2016)

2.

One pipeline, four cognitive functions
→
Every request goes through the same steps. Operators use the same chain, ranked by risk instead
Input
free text
live sensors
Perception
› entities, features, ›
forecast
Attention
score 30 car
parks, keep ~6
Memory
›
this session +
SQLite database
Knowledge
›
graph + rules
R0–R10
Output
› recommendation,
warning, advice
Data
Two users
Real: UCI Parking Birmingham. 30 car parks, readings
every 30 min, 08:00–16:30, Oct–Dec 2016
Knowledge base: 10 zones, 17 landmarks, facilities,
tariffs. An illustrative enrichment (JSON)
Driver: "electric car to the Arena Friday 5pm, 4 hours,
cheapest" gets a car park, cost, free-space forecast and
reasons
Operator: "which car parks are at risk?" gets a priority
ranking, anomaly alerts and diversion advice
Stack: pandas · scikit-learn · networkx · SQLite

3.

1
Perception: from raw signals to structured facts
Two inputs: live sensor counts and the driver's own words
0.069
forecast error (MAE) on 14 held-out days
vs 0.240 naive · 3.5× better
Cleaning: 216 duplicates dropped, 12 negative counts clipped, 373 overcapacity flagged, timestamps snapped to 30-min slots
Features: how full, time slot, weekday, recent change, usual level, difference
from usual
Language: word lists, patterns and typo matching. "bulring tomorow 11am" →
Bullring, Saturday, 11:00
Why word lists, not a model?
Requests are short and there is no labelled
parking data to train on. Every extracted
detail can be explained, and anything missing
is filled in from memory.

4.

2
Attention: focus on ~6 of 30 car parks
score = .20 relevance + .35 proximity + .35 availability + .10 reliability → softmax → threshold 1/N
HIGH
LOW
Car park
Bullring CP 3
CPS105a
Zone hops
0
3
TF-IDF relevance
1.00
0.00
Free at peak
38 %
25 %
Attention weight
0.323
0.002
Selected
yes
no
Operator attention
Risk = projected fill in 1 h. On Sat 17 Dec at 13:00:
Town Hall 100 % CRITICAL vs Europark BS2 24 %
LOW. An Isolation Forest marks about 2 % of readings
as strange, e.g. an event at the arena.

5.

3
Memory: context now, experience later
Short-term memory understands follow-ups; long-term memory changes later decisions
Short-term
Long-term (SQLite)
Same request, after restart
last 10 events and the current
request, kept for one session
users, requests, suggestions +
feedback, alerts
"What about somewhere else?"
reuses the destination, time and
budget, and skips car parks already
suggested
stays after the program closes
read back: vehicle, past feedback,
repeat hot-spots
"Broad Street, Monday 9am, all day"
no memory → NCP Pershore St
(no EV charger!)
with memory → Europark BS 1
remembers the EV; lowers the car park
they rejected
✓
Licence plate
Privacy by design
✓
Accessibility need
health data: stored only with consent
salted SHA-256 hash, never clear text
✓
History
zone only, no GPS · forget_user() erases

6.

4
Knowledge: a graph of facts plus if-then rules
70 nodes · 197 edges · 6 entity types · 5 relations · rules R0–R10
R1–R3
HARD EV needs a charger · accessible
R4–R5
SOFT tariff fits stay length · budget
R6–R7
WARN predicted ≥ 90 % full · unreliable
R8–R9
SOFT not open 24 h · personal history
R0
INFER widen the search 1 zone hop at a
R10
OPERATOR divert CRITICAL traffic to a
Why this car park: Utilita Arena → in Arena zone → next to Broad Street → Europark Broad Street 1 is there
bay · van height limit
bonus
sensor
(LTM)
time
calm neighbour

7.

Live demo & results
Takeaways
Input
Output
Used
Bullring, Sat 1pm, 3 h
Bullring CP 3 · 38 % free
R4 R6 R7
EV → Arena, Fri 5pm
NIA South (charger)
R1 R0 R5 R8
"somewhere else?"
Europark BS 1
memory, R0 R1
Broad St, Mon, all day (restart)
Europark BS 1 + 91 % warning
memory, R6 R9
Operator, Sat 17 Dec 13:00
5 CRITICAL → divert
R10 + alerts
Van near the markets
CPS119a (no height limit)
R3 → R0
Run it:
jupyter notebook AI_Parking_Management_System.ipynb
·
Every answer comes with its reasons:
score parts, rules used, graph path
Memory changes the answer, not just
the log
Limits: forecast uses past averages only,
sensors cover 08:00–16:30, weights set
by hand, facilities made up
Next: learn weights from feedback, add
event calendars, walking distance,
camera perception
python main.py --demo